Designing Network Protocols(journal.paul.querna.org) |
Designing Network Protocols(journal.paul.querna.org) |
Unfortunately serviceability is not normally high in the initial product requirements, but I've seen a direct correlation between customer satisfaction in support and products/protocols/designs with good serviceability.
Some people just like ASCII, human readable protocols. There's nothing wrong with that, but it's a little silly to suggest that the options for a packed binary encoding in 2007 were limited because Thrift and Protocol Buffers were too new.
I'd be interested in the original designer's remarks on using existing wire serialisation formats.
I enjoyed reading through his thought process for designing a simple protocol which is:
* Easy to use (requiring little/no additional libraries)
* Easy to extend (simple keyword/value extensions)
* Immune to changes in technology
* (above all) easy to understandThat said, I think he didn't want a binary format in any case -- his "doing it again today" remarks point to JSON.
http://catb.org/~esr/writings/taoup/html/textualitychapter.htmlIt's obviously doable, but it's very painful.
I can see both sides of the argument here, but basing a protocol on text just for the ease of eyeballing it on-the-wire seems like optimizing for the uncommon case.
Heck, almost any decent protocol should only have ciphertext on-the-wire anyway.
I get the rationale. But I think it's weak, and this entire post is lots of fluff around that core rationale. (I've been writing extensible binary protocols back in 1988 - and it never struck me as particularly difficult even back then.)
But even still, this only matters if:
A. The protocol is so new that Wireshark isn't shipping a parser,
B. the admin's stuff isn't working,
C. the admin can't get his stuff working by normal troubleshooting and must resort to observing the protocol,
D. the admin can't get his stuff working by observing the binary representation of the protocol, and
E. the admin actually can get his stuff working with a transliterated ASCII representation of the protocol.
Certainly I would probably find it easier to troubleshoot a text-based protocol too. I just think it's a relatively minor case in the grand scheme of things.
My point is that I imagine a network designer shouldn't focus on Wireshark or tcpdump integration over other non-functional requirements such as, well, network performance.
Network performance isn't as visible as the non-functional requirement of inspectability because it is amortised over potentially millions of machines, whereas inspectability is an immediately visible issue to the select few who "pop the hood" to fix an issue or simply to have a look.
For example: in terms of network capacity, I wonder how much HTTP headers cost all of us collectively. Probably a lot more than the cost of making a Wireshark plugin and having sysadmins install it as necessary.
Edit: put another way, I think designers should prioritise the needs of the people who pay the cost of network operation over the convenience of the operators.
There's a feedback loop here -- if it's too hard and thus very expensive to operate a system, then optimising for performance was a false win. But I don't think this is such a case, especially since as you pointed out elsewhere there are a number of very mature binary wire formats that were extant in 2007.
"I implore [you] to remember Dave and Virginia, preying on the drug addicts of the next generation and the sexually dissatisfied men of the previous generation. How different their careers could have been if their parents had not downloaded so many terabytes of data! We must not abandon our children to such a fate."